Computer vision development
A computer vision system turns pixels into a result another application can use. That might be a label or a count, an object's location or outline, a measurement, a field read from a document or an alert.
Model accuracy is only one factor in whether it works in production. Outside the test dataset, cameras, lighting and motion change what the model sees. Labeling rules, decision thresholds, review steps and the hardware it runs on shape what it does with it.
What we build
The output your application needs defines the task, and one image can feed several tasks at once.
Classification
Assign one or more labels to an image or a cropped region. Sorting document pages by type is a typical use.
label + confidence
Object detection
Find and classify each relevant object in a frame, such as items to count or damaged packages in a warehouse.
boxes + labels
Segmentation
Mark the pixels that belong to an object, surface or condition.
masks + area
Tracking
Follow the same object through video and record its movement or state, such as a loading event or a defined condition on site.
track + events
OCR and extraction
Read text and map fields from documents, labels, screens or scenes, keeping page coordinates for review.
text + fields
Visual inspection
Find defects, missing parts or process deviations on a production line, and send unclear cases to an operator.
finding + location
When it's a good fit
Custom training pays off
The target, the environment or the error tolerance is specific enough that general models are unreliable.
A ready-made model may be enough
If a pretrained model or a managed vision API meets the requirements, we use it instead of training a new one.
Multimodal may fit better
When the visual result has to be compared with a manual, a specification or a spoken note, the task needs more than the image and becomes multimodal work.
How a project runs
-
Look at your cameras and footage
Before choosing a model, we check sample footage to see whether image quality supports the task and which conditions your data is missing.
-
Agree on labeling rules
The guidelines spell out what counts as an object, defect, boundary or event. We first have more than one person label the same sample. Where they disagree, the rule is unclear, and we fix it before the full dataset is labeled.
-
Set thresholds by the cost of an error
Missing a safety event and sending one extra item for review do not cost the same. We set the confidence and duration rules that create, suppress or route an alert with that in mind. Unclear cases can also go to a person.
-
Trial it on site
During a monitored trial we log missed cases and false alerts, note camera changes and collect operator feedback. We also measure the time from frame to result and the load on the hardware.
The test set has to look like your site
A model can score well on clean images and still fail after installation. The evaluation data covers the conditions below, and we report results for each site and condition.
Light
Day, night, shadows, reflections, glare and seasonal change.
View
Camera height, angle, distance, lens, crop and resolution.
Scene
Objects hiding each other, clutter, look-alike objects, motion blur and changing backgrounds.
Case mix
Common cases, rare faults, new variants and long periods with nothing to detect.
Hardware
Camera compression, frame rate, connectivity, compute and installation limits.
Privacy and responsibility
Privacy zones, retention, access, review responsibilities and lawful use.
What you receive
- Task and camera feasibility assessment
- Labeling guide and versioned datasets
- Model or vision API integration, running in the agreed environment
- Evaluation set, thresholds, retraining triggers and test report
- Code and documentation, handed over to your team
Frequently asked questions
Can we use our existing cameras for computer vision?
Often, yes. The footage check at the start shows whether they capture what needs to be detected clearly enough.
Do we need to label our own images?
In most cases. Ready-made models recognize common objects, but your own products or defects need labeled examples. The first pilot shows how many more are worth collecting.
Can computer vision work in real time?
It can, if the model and hardware are chosen with that in mind. Not every task needs every frame. Some only need to look when something happens.
Where are camera images processed?
That depends on the site. It can be the cloud, your own infrastructure, or a device next to the camera when latency, bandwidth or offline operation matter. Images can stay on your premises, with only selected events sent for central review.
Technical references
Our products in this area
Work with us
Tell us in a few lines what you're working on. We'll get back to you.
From our blog
All articles
What AI evaluation metrics miss in production
Strong AI evaluation scores can hide production failures when the chosen metric does not reflect business risk, user outcomes, or operational cost.
Read article
Will a proven AI method work on your data?
A proven AI method can still fail on your data and workflow. Testing shows whether published evidence transfers to your specific constraints.
Read article
When should you use synthetic data?
Synthetic data can help when real data is scarce, sensitive, or unavailable. When is it worth using, and how can you tell whether it improves results?
Read article